Abstract: For cascade reservoirs in the period after main flood process, decision-making on the timing, rate and sequence of their refilling is faced with problems such as runoff forecast uncertainty and complex multi-objective competition-cooperation relationships. To deal with such difficulties in the refilling decision-making, we develop a stochastic optimization operation model for the water refilling of cascade reservoirs with five objectives, power generation, refilling degree, ecology, and upstream and downstream flood risks. We generate a non-inferior solution set for the refiling, use the K-means clustering method to extract its features, and analyze the contradiction relationship between the objectives and the refilling mechanism. Application to a case study of the cascade reservoirs on the Yangtze River mainstream from Xiluodu to Gezhouba shows a drier water refilling period leads to a greater difficulty in the refilling and lower comprehensive benefits. For these reservoirs, the total power output in a dry year is 21.6% lower than that of a wet year; the strongest contradiction occurs between upstream flood control safety and power generation, with the correlation coefficient decreasing by 0.047 from a dry to wet year. The refilling in the Xiluodu and Three Gorges reservoirs varies in stages, rapid in the early stage and slow in the middle and later stages. The multi-objective stochastic programming and clustering analysis method developed in this study for reservoir refilling helps formulate better refilling schemes.

Key words: reservoir operation, cascade reservoirs, multi-objective optimization, joint operation, K-means clustering